A Cluster Photovoltaic Power Prediction Method and Device Based on Ground-Based Cloud Images

Abstract

The present invention proposes a method and device for predicting the power of a cluster of photovoltaic power stations based on ground-based cloud images, which record the location information of each photovoltaic power station in the cluster, collect the ground-based cloud image data and power data of each photovoltaic power station, and obtain numerical weather prediction (NWP) data; preprocess the NWP data, and use the maximum relevance minimum redundancy algorithm to screen out meteorological factors with a relatively high degree of correlation to obtain an optimized meteorological feature data set; preprocess the ground-based cloud image data, perform registration and fusion operations on the ground-based cloud images at different points, extract the key features of the cloud images to obtain a cloud image key feature data set; train a cluster photovoltaic power prediction model based on the optimized meteorological feature data set and the cloud image key feature data set, and obtain the predicted power of the cluster of photovoltaic power stations according to the model. The present invention realizes high-precision feature extraction and image fusion, optimizes the prediction of the power of a cluster of photovoltaic power stations, and overcomes the problem in the prior art that the ultra-short-term photovoltaic power of the cluster cannot be predicted comprehensively, timely and accurately.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy power, and particularly relates to a method and device for predicting the power of a cluster of photovoltaic power based on ground-based cloud images. Background Art

[0002] As a clean and renewable energy source, solar energy has been increasingly widely concerned by people. With the continuous expansion of the scale of photovoltaic grid connection, photovoltaic has become the third largest power energy. However, the intensity of sunlight received on the ground is greatly affected by weather conditions, cloud movement, etc. When the penetration power of photovoltaic is high, it will cause an impact on the power grid. Therefore, in order to ensure the safe and reliable operation of the power grid and reduce the phenomenon of "light abandonment and power curtailment", it is very important to accurately predict the irradiance and photovoltaic power.

[0003] Currently, the commonly used methods for predicting the power of a photovoltaic cluster mainly include the cumulative method and the statistical upscaling method. The cumulative method obtains the power prediction of the photovoltaic cluster by directly adding up the prediction results of all photovoltaic power stations in the cluster, while the statistical upscaling method calculates the power prediction of the photovoltaic cluster through upscaling using the prediction results of the reference photovoltaic power station. These methods have deficiencies in the in-depth mining of the input data for power prediction, and do not consider the temporal relationship and spatial geographical location similarity among multiple photovoltaic power stations, resulting in the need to improve the prediction accuracy.

[0004] In addition, the data used for ultra-short-term photovoltaic forecasting of the cluster mainly include numerical weather forecasting and historical power generation. Numerical weather forecasting provides specific values of various meteorological elements, which can significantly reduce the systematic error, but has a low spatio-temporal resolution, a large spatio-temporal scale, and does not have the ability to improve the forecasting of intraday cloudy days, sudden weather changes, etc., and cannot reflect the details of weather changes at a higher temporal resolution, and the update frequency is generally once every 12 hours. It cannot meet the high-precision ultra-short-term prediction requirements of each power station. If minute-level ground-based cloud image data is used, a more accurate estimate of the forecasting error can be given. However, the observation range of the ground-based cloud image only covers the sky information over the local real-time area, so it is mostly used for single power station power prediction. Since the photovoltaic power stations in the cluster may be located in different geographical locations, there are differences in the spatial distribution of the ground-based cloud image data. How to effectively fuse it and use it for cluster power prediction is still a challenge. Summary of the Invention

[0005] Aiming at the problems in the prior art, the present invention proposes a method and device for predicting the power of a cluster of photovoltaic power based on ground-based cloud images, which realizes high-precision feature extraction and image fusion, optimizes the prediction of the power of the cluster of photovoltaic power, and overcomes the problem that the prior art cannot comprehensively, timely, and accurately perform ultra-short-term photovoltaic power prediction for the cluster.

[0006] To achieve the above object, the technical solution of the present invention is realized as follows:

[0007] A method for predicting the power of a cluster of photovoltaic power plants based on ground-based cloud images, comprising:

[0008] S1. Data collection: Record the location information of each photovoltaic power plant in the cluster, collect the ground-based cloud image data and power data of each photovoltaic power plant, and obtain numerical weather prediction (NWP) data;

[0009] S2. Preprocessing of NWP data: Use the maximum correlation minimum redundancy algorithm to screen out meteorological factors with a relatively high degree of correlation, and obtain an optimized meteorological feature data set;

[0010] S3. Preprocessing of ground-based cloud image data: Perform registration and fusion operations on ground-based cloud images at different points, extract key features of the cloud images, and obtain a key feature data set of the cloud images;

[0011] S4. Train a cluster photovoltaic power prediction model based on the optimized meteorological feature data set and the key feature data set of the cloud images, and obtain the predicted power of the cluster photovoltaic power according to the model.

[0012] Furthermore, in step S1, obtaining numerical weather prediction (NWP) data includes obtaining numerical weather prediction data sets from different meteorological sources in the same time period. The data sets include irradiance, wind speed, wind direction, temperature, humidity, and air pressure data; process the abnormal values in the irradiance data that violate the true light law.

[0013] Furthermore, step S2 includes:

[0014] S201. Perform maximum-minimum normalization processing on the obtained NWP data;

[0015] S202. Use the maximum correlation minimum redundancy algorithm, with the power data of the cluster photovoltaic power as the target variable, calculate the correlation between the NWP data of different meteorological sources and the target variable, and obtain an optimized meteorological feature data set by maximizing the correlation and simultaneously minimizing the redundancy between the NWP data.

[0016] Furthermore, step S3 includes:

[0017] S301. For the ground-based cloud image data of each power plant, retain the smallest circumscribed rectangle containing the spherical mirror, and use median filtering to process the noise points to obtain the preprocessed ground-based cloud image;

[0018] S302. Calculate the distances between the photovoltaic power plants according to the longitude and latitude data of the location information of each photovoltaic power plant, and store them in an adjacency matrix;

[0019] S303. Sort according to the distances between the photovoltaic power plants to obtain the distance relationship of the points where each power plant is located. Perform image registration on the preprocessed ground-based cloud images in order from near to far, and fuse the registered images;

[0020] S304. Extract key cloud map features from the fused ground-based cloud map fused image to obtain a key cloud map feature dataset; the key cloud map features include light intensity, cloud cover percentage, transmittance, and zenith distance.

[0021] Furthermore, step S4 includes:

[0022] S401. Treat each photovoltaic power station as a node, represent the relationship between nodes through the adjacency matrix, and for the preferred meteorological feature dataset and the key cloud map feature dataset, use a graph attention network to achieve node-level information aggregation and feature embedding to obtain the spatial features of each node.

[0023] S402. Based on the Scaleformer network model, perform multi-scale processing, use the spatial features of each node as input variables, use the power data of the cluster photovoltaic as the target variable, and aim to minimize the loss function, and obtain a cluster photovoltaic power prediction model through training.

[0024] S403. Make a prediction according to the cluster photovoltaic power prediction model to obtain the final cluster power prediction value.

[0025] On the other hand, the present invention also proposes a cluster photovoltaic power prediction device based on ground-based cloud maps, including:

[0026] Data collection module: Record the location information of each photovoltaic power station in the cluster, collect the ground-based cloud map data and power data of each photovoltaic power station, and obtain numerical weather prediction NWP data.

[0027] Meteorological feature module: Used for preprocessing NWP data, and use the maximum correlation minimum redundancy algorithm to screen out meteorological factors with a relatively high degree of correlation to obtain a preferred meteorological feature dataset.

[0028] Cloud map feature module: Used for preprocessing ground-based cloud map data, perform registration and fusion operations on ground-based cloud maps at different points, extract key cloud map features, and obtain a key cloud map feature dataset.

[0029] Prediction model module: Train a cluster photovoltaic power prediction model based on the preferred meteorological feature dataset and the key cloud map feature dataset, and obtain the predicted power of the cluster photovoltaic according to the model.

[0030] Further, in the data collection module, obtaining numerical weather prediction NWP data includes obtaining numerical weather prediction datasets from different meteorological sources in the same time period, and the datasets include irradiance, wind speed, wind direction, temperature, humidity, and air pressure data; process the outliers in the irradiance data that violate the true light law.

[0031] Further, the meteorological feature module includes:

[0032] Normalization unit: Perform maximum-minimum normalization on the obtained NWP data;

[0033] Correlation unit: Use the maximum correlation and minimum redundancy algorithm, with the power data of the cluster photovoltaic as the target variable, calculate the correlation between the NWP data of different meteorological sources and the target variable, and obtain the optimal meteorological feature dataset by maximizing the correlation and minimizing the redundancy between the NWP data at the same time.

[0034] Furthermore, the cloud map feature module includes:

[0035] Preprocessing unit: For the ground-based cloud map data of each station, retain the smallest circumscribed rectangle containing the spherical mirror, and use median filtering to process the noise points to obtain the preprocessed ground-based cloud map;

[0036] Matrix unit: Calculate the distances between the photovoltaic stations according to the latitude and longitude data of the position information of each photovoltaic station, and store them in an adjacency matrix;

[0037] Fusion unit: Sort according to the distances between the photovoltaic stations to obtain the distance relationship of the positions of each station. Perform image registration on the preprocessed ground-based cloud maps in order from near to far, and fuse the registered images;

[0038] Feature extraction unit: Extract the key features of the cloud map from the fused ground-based cloud map fusion image to obtain the key feature dataset of the cloud map; The key features of the cloud map include light intensity, cloud cover percentage, transmittance, and zenith distance.

[0039] Even further, the prediction model module includes:

[0040] Spatial feature unit: Regard each photovoltaic station as a node, represent the relationship between nodes through the adjacency matrix, and for the optimal meteorological feature dataset and the key feature dataset of the cloud map, use a graph attention network to achieve node-level information aggregation and feature embedding to obtain the spatial features of each node;

[0041] Model unit: Based on the Scaleformer network model, adopt multi-scale processing, use the spatial features of each node as the input variable, use the power data of the cluster photovoltaic as the target variable, and aim to minimize the loss function, and obtain the cluster photovoltaic power prediction model through training;

[0042] Prediction unit: Make a prediction according to the cluster photovoltaic power prediction model to obtain the final cluster power prediction value.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] In view of the situation that each photovoltaic power station located in different geographical positions in the cluster photovoltaic power prediction has different spatio-temporal characteristics, through high-precision feature extraction and image fusion, the ground-based cloud map data of each power station is effectively fused, the sky image information in different spaces is fully utilized, and more reliable and comprehensive regional cloud map observation results are provided, so as to provide more accurate ultra-short-term cluster photovoltaic power prediction, reduce operation and maintenance costs, and enhance power grid stability. Brief Description of the Drawings

[0045] Figure 1 It is a schematic flowchart of an embodiment of the present invention.

[0046] Figure 2 It is a schematic diagram of the spatial distribution of a certain cluster photovoltaic power station in an embodiment of the present invention. Detailed Embodiments

[0047] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0048] To make the purpose and features of the present invention patent more obvious and understandable, the present invention will be further described below in conjunction with the drawings and specific embodiments.

[0049] The cluster photovoltaic power prediction method based on ground-based cloud maps proposed by the present invention is as Figure 1 shown, and includes the following steps:

[0050] S1. Data collection: Record the location information of all photovoltaic power stations in the cluster, collect the ground-based cloud map data of each photovoltaic power station by using a sky imager, obtain the power data of each photovoltaic power station through a supervisory control and data acquisition system, and obtain numerical weather prediction (NWP) data;

[0051] S2. NWP data preprocessing: Use the maximum relevance minimum redundancy algorithm to screen out meteorological factors with relatively high correlation degrees to obtain an optimized meteorological feature data set;

[0052] S3. Ground-based cloud map data preprocessing: Perform registration and fusion operations on the ground-based cloud maps of different power stations, extract the key features of the cloud maps, and obtain a cloud map key feature data set;

[0053] S4. Train a cluster photovoltaic power prediction model based on the optimized meteorological feature data set and the cloud map key feature data set, and obtain the cluster photovoltaic predicted power according to the model.

[0054] In the step S1, the cluster photovoltaic consists of multiple photovoltaic power stations distributed in different spaces, as Figure 2 shown, where 1, 2, 3, 4, 5, and 6 respectively represent photovoltaic power stations in different spaces. Therefore, the data collection includes the following steps:

[0055] (1-1) Record the location information of each photovoltaic power station; prepare the historical observed power of each photovoltaic power station to obtain a historical power dataset;

[0056] (1-2) Prepare a numerical weather prediction (NWP) dataset of different meteorological sources for the same time period as the historical power dataset. The data in the NWP dataset includes six meteorological characteristics: irradiance, wind speed, wind direction, temperature, humidity, and air pressure.

[0057] (1-3) Process the outliers in the irradiance data of the NWP dataset that violate the true lighting law. The violations of the true lighting law include two cases: the night value is not 0 and the day value does not change for two consecutive hours. Delete the outliers that meet these two cases, and calculate the average value of the irradiance data values at the same time of the previous and next seven days, and replace the outliers with the average value.

[0058] (1-4) Obtain the ground-based cloud map data of each photovoltaic power station collected by the sky imager for the same time period as the historical dataset.

[0059] The step S2 includes the following steps:

[0060] (2-1) Perform maximum-minimum normalization on the dataset of numerical weather prediction (NWP) data of different meteorological sources to eliminate the influence of different dimensions.

[0061] ;

[0062] where is the k-th value of the j-th variable of the i-th station in the time series of the NWP dataset after maximum-minimum normalization, is the k-th value of the j-th variable of the i-th station in the time series of the original NWP dataset, and are the minimum and maximum values of the j-th variable of the i-th station in the time series of the original NWP dataset, respectively.

[0063] (2-2) Use the maximum correlation minimum redundancy algorithm to calculate the meteorological characteristics in the NWP dataset from different sources and the cluster photovoltaic power value. The cluster photovoltaic power value is the sum of the observed powers of all photovoltaic power stations in the cluster for the same time period in the power dataset. Take the cluster photovoltaic power value as the target variable, and by maximizing the correlation between the meteorological characteristics in the NWP dataset and the target variable, and minimizing the redundancy between the meteorological characteristics in the NWP dataset at the same time, improve the generalization ability of the model and reduce the risk of overfitting to obtain an optimized meteorological characteristic dataset.

[0064] The step S3 includes the following steps:

[0065] (3-1) Preprocess the ground-based cloud map data of each photovoltaic power station collected by the sky imager. Crop the unnecessary information around and only retain the smallest circumscribed rectangle containing the spherical mirror. In the acquisition and processing flow of the ground-based cloud map data, since the all-sky imager is affected by the disturbance of the transmission medium or the defects of electronic devices during image acquisition, and it is inevitable to be disturbed by various noises during the shooting, recording, and transmission processes, which usually appear as individual pulse signals on the image and as a pixel block or isolated pixel points on the graph, affecting the target display as useless information. The basic principle of the median filter is a non-linear signal processing method designed based on the sorting statistics theory, which can be effectively applied in the case of impulse noise and salt-and-pepper noise and can better retain the texture information and edge details of the image. Therefore, the median filtering method is used to solve some of the noise points in the ground-based cloud map images.

[0066] ;

[0067] Among them is the pixel value of the filtered image at position , is the pixel value of the original image at position , and are the indices of the pixels in the traversed neighborhood, represents the median function, taking the median of all .

[0068] (3-2) According to the location information of each photovoltaic power station, extract the longitude and latitude data, calculate the distances between each pair, and use the Euclidean distance to represent the distance relationship in this embodiment; after calculating the Euclidean distances between each photovoltaic power station, store the calculation results in an adjacency matrix; the construction of the adjacency matrix takes each photovoltaic power station as a node and uses the Euclidean distances between each photovoltaic power station to represent the edges between the nodes, and it is constructed based on the KNN algorithm.

[0069] (3-3) Sort the distance values of each station in the cluster to obtain the proximity relationship of each node (each photovoltaic power station), and perform image registration in order from near to far, and then fuse the registered images;

[0070] Among them, step (3-3) includes the following sub-steps:

[0071] (3-3-a) Sort the distance values (Euclidean distances) of each station in the cluster, and first select the ground-based cloud map images of the 2 stations with the closest distance;

[0072] (3-3-b) Combine the efficiency of SURF (Speeded Up Robust Features) feature extraction with the rotational invariance and robustness of the ORB (Oriented FAST and Rotated BRIEF) descriptor for image registration of the ground-based cloud images of the two nearest stations. Detect feature points and generate descriptors through SURF (Speeded Up Robust Features), and then use the ORB algorithm for fast matching. After image registration, each image in the image sequence needs to be projected and transformed to the reference image according to the transformation matrix between the images to complete panoramic stitching. Use the traditional bundle adjustment method to optimize the transformation matrix to improve the accuracy of the transformation matrix and obtain the stitched image of the two images;

[0073] (3-3-c) In the ground-based cloud images of the remaining PV stations, successively take out the nearest 1 image and stitch it with the already fused image, repeating the process of step (3-3-b) until the ground-based cloud images of all stations are completely stitched to obtain the fused ground-based cloud image;

[0074] The above image registration algorithm can also be replaced by its homologous algorithms such as the SIFT algorithm.

[0075] (3-4) Extract key features such as light intensity, cloud cover percentage, transmittance, and zenith distance from the fused ground-based cloud images of each time period to obtain the key feature dataset of the cloud images;

[0076] Among them, step (3-4) includes the following sub-steps:

[0077] (3-4-a) The fused ground-based cloud image is an RGB image. Convert the RGB image to the color space model HSV to obtain the light intensity feature. The calculation formulas for each component of the actual ground-based cloud AHSV model are as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] In the formula, AR, AG, and AB are the red, green, and blue color channel values in the actual ground-based cloud ARGB model; max = max(AR, AG, AB), representing the largest of the three color channel values; min = min(AR, AG, AB), representing the smallest of the three color channel values; AH is the hue component value, characterizing the color of the image, that is, the position information of the spectral color, expressed in degrees, where 0 o 、120 o 、240 oThey respectively correspond to red, green, and blue; AS is the saturation component value, representing the numerical feature of color depth; AV is the lightness component value, that is, the color brightness level feature, and its value corresponds to the interval [0, 1], indicating that the value range is 0.0 (black) to 1.0 (white).

[0082] The calculated AH, AS, and AV are used as the illumination intensity features of the ground-based cloud image fusion image.

[0083] (3-4-b) Extract the key feature of the sky cloud cover from the ground-based cloud image fusion image;

[0084] ;

[0085] Among them, F represents the sky cloud cover; represents the number of cloud pixels, represents the number of pixels in the sky area;

[0086] (3-4-c) The transmittance will affect the solar energy received by the photovoltaic panel and can reflect the clarity of the ground-based cloud image. The actual ground-based cloud image transmittance is calculated as follows:

[0087] ;

[0088] In the formula, is the estimated value of the actual cloud image transmittance, J C is a certain color channel of J, is a square area, and its center point is x a , and A is the global sunlight component;

[0089] (3-4-d) The zenith distance reflects the relative distance between the sun and the all-sky imager. Then the zenith distances Ad of the actual cloud image and the predicted cloud image are as follows:

[0090] ;

[0091] In the formula, (x 0 , y 0 ) is the position corresponding to the all-sky imager in the ground-based cloud image, and (x A , y A ) is the position corresponding to the sun in the actual ground-based cloud image.

[0092] The step S4 includes the following steps:

[0093] (4-1) Each photovoltaic power station is used as a node of the adjacency matrix, and the distance between photovoltaic power stations represents the edge between nodes. For the preferred meteorological features in the preferred meteorological feature dataset and the fused cloud image features, a graph attention network is used to achieve node-level information aggregation and feature embedding to capture the potential spatial structure information in the data.

[0094] Among them, the input of the graph attention network is the location relationship between photovoltaic power stations and the characteristic data of each power station itself. By introducing the attention mechanism to aggregate the node information of different power stations, for the neighbor power station information around each power station in the adjacency matrix, weighted aggregation is performed using the weights learned by the model. First, a linear transformation is performed on the node features, then the attention scores between nodes are calculated using the attention mechanism and normalized. Finally, the power station information around each power station is weighted and aggregated to obtain high-dimensional features that fuse the information of surrounding power stations.

[0095] The graph attention network uses the self-attention mechanism to calculate the attention weights between each photovoltaic power station node in the adjacency matrix and its neighboring power station nodes, so that the state update of each photovoltaic power station node will consider the states of its neighboring power station nodes; the features of the nodes are aggregated by weighted summation, and the weights are determined by the attention coefficients. In this way, the model can automatically learn which neighboring power station nodes around the photovoltaic power station are more important for the state update of the current node; the multi-head attention mechanism is adopted, and each attention head independently calculates the attention scores and finally concatenates the results of all heads, enabling the model to focus on different features or patterns and thus capture more information. Through these characteristics, the graph attention network can effectively capture the potential spatial structure information in the cluster photovoltaic data and provide strong support for the graph data analysis task of cluster photovoltaic power prediction.

[0096] (4-2) The Scaleformer network model adopts a multi-scale processing method. Based on the spatial features and target cluster power in the previous step, with the goal of minimizing the loss function, it is obtained through training to generate the final cluster power prediction value. While ensuring high accuracy by adopting multi-scale, the computational overhead is reduced by sharing weights and optimizing the architecture design, which is beneficial to improving the accuracy, real-time performance, and robustness of the prediction.

[0097] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A cluster photovoltaic power prediction method based on ground-based cloud map, characterized in that: include: S1. Data collection: record the location information of each photovoltaic station in the cluster, collect ground-based cloud map data and power data of each photovoltaic station, and obtain numerical weather forecast NWP data; S2, NWP data preprocessing, using the maximum correlation minimum redundancy algorithm to screen out meteorological factors with high correlation and obtain the optimal meteorological characteristic data set; S3, preprocessing of ground-based cloud map data, performing registration and fusion operations on ground-based cloud maps at different points, extracting key features of cloud maps, and obtaining cloud map key feature data sets; S4, training a cluster photovoltaic power prediction model based on the preferred meteorological feature data set and the cloud image key feature data set, and obtaining cluster photovoltaic predicted power according to the model; Step S3 includes: S301, retaining the minimum circumscribed rectangle containing the spherical mirror for the ground-based cloud image data of each station, using median filtering to process noise points, and obtaining a pre-processed ground-based cloud image; S302, calculating the distance between photovoltaic stations according to the longitude and latitude data of the location information of each photovoltaic station, and storing it in an adjacency matrix; S303, sorting the photovoltaic stations according to the distances between them to obtain the distance relationship between the locations of the stations, performing image registration of the pre-processed ground cloud images in order from near to far, and fusing the registered images; S304, extracting cloud map key features from the fused ground-based cloud map fusion image to obtain a cloud map key feature data set; the cloud map key features include light intensity, cloud cover percentage, transmittance and zenith distance; Step S4 includes: S401, treating each photovoltaic station as a node, representing the relationship between nodes through the adjacency matrix, and using the graph attention network to achieve node-level information aggregation and feature embedding for the optimal meteorological feature data set and the cloud map key feature data set, to obtain the spatial features of each node; S402, based on the Scaleformer network model, multi-scale processing is adopted, the spatial characteristics of each node are used as input variables, the power data of the cluster photovoltaic is used as the target variable, and the minimum loss function is used as the goal, and the cluster photovoltaic power prediction model is obtained through training; S403: Predicting according to the cluster photovoltaic power prediction model to obtain a final cluster power prediction value.

2. The cluster photovoltaic power prediction method based on ground-based cloud map according to claim 1 is characterized in that: In step S1, obtaining numerical weather forecast NWP data includes obtaining numerical weather forecast data sets from different meteorological sources in the same time period, the data sets including irradiance, wind speed, wind direction, temperature, humidity, and air pressure data; and processing abnormal values ​​in irradiance data that violate the actual lighting laws.

3. The cluster photovoltaic power prediction method based on ground-based cloud map according to claim 1 is characterized in that: Step S2 includes: S201, performing maximum and minimum normalization processing on the acquired NWP data; S202, using the maximum correlation minimum redundancy algorithm, taking the power data of the cluster photovoltaic as the target variable, calculating the correlation between the NWP data of different meteorological sources and the target variable, and obtaining the optimal meteorological characteristic data set by maximizing the correlation and minimizing the redundancy between the NWP data.

4. A cluster photovoltaic power prediction device based on ground-based cloud map, characterized in that: include: Data collection module: records the location information of each photovoltaic station in the cluster, collects ground-based cloud map data and power data of each photovoltaic station, and obtains numerical weather forecast NWP data; Meteorological feature module: used for NWP data preprocessing, using the maximum correlation minimum redundancy algorithm to screen out meteorological factors with high correlation and obtain the optimal meteorological feature data set; Cloud map feature module: used for ground-based cloud map data preprocessing, registering and fusing ground-based cloud maps at different points, extracting key features of cloud maps, and obtaining cloud map key feature data sets; Prediction model module: training a cluster photovoltaic power prediction model based on the preferred meteorological feature data set and the cloud image key feature data set, and obtaining cluster photovoltaic predicted power according to the model; The cloud map feature modules include: Preprocessing unit: For the ground-based cloud image data of each station, the minimum circumscribed rectangle containing the spherical mirror is retained, and the noise points are processed using median filtering to obtain the preprocessed ground-based cloud image; Matrix unit: Calculate the distance between photovoltaic stations based on the longitude and latitude data of the location information of each photovoltaic station and store it in an adjacency matrix; Fusion unit: Sort the photovoltaic stations according to their distances to obtain the distance relationship between the locations of the stations, perform image registration of the pre-processed ground cloud images from near to far, and fuse the registered images; Feature extraction unit: extract key features of cloud images from the fused ground-based cloud image to obtain a cloud image key feature data set; the key features of cloud images include light intensity, cloud percentage, transmittance and zenith distance; The prediction model module includes: Spatial feature unit: each photovoltaic station is regarded as a node, and the relationship between nodes is represented by the adjacency matrix. For the optimal meteorological feature data set and cloud image key feature data set, the graph attention network is used to realize node-level information aggregation and feature embedding to obtain the spatial features of each node; Model unit: Based on the Scaleformer network model, multi-scale processing is adopted, the spatial characteristics of each node are used as input variables, the power data of cluster photovoltaics is used as the target variable, and the minimum loss function is used as the goal. The cluster photovoltaic power prediction model is obtained through training; Prediction unit: performs prediction according to the cluster photovoltaic power prediction model to obtain a final cluster power prediction value.

5. The cluster photovoltaic power prediction device based on ground-based cloud map according to claim 4 is characterized in that: In the data collection module, obtaining numerical weather forecast NWP data includes obtaining numerical weather forecast data sets from different meteorological sources in the same time period, which include irradiance, wind speed, wind direction, temperature, humidity, and air pressure data; and processing abnormal values ​​in irradiance data that violate the actual lighting laws.

6. The cluster photovoltaic power prediction device based on ground-based cloud map according to claim 4, characterized in that: Meteorological feature modules include: Normalization unit: perform maximum and minimum normalization processing on the acquired NWP data; Correlation unit: Using the maximum correlation minimum redundancy algorithm, with the power data of cluster photovoltaics as the target variable, the correlation between the NWP data of different meteorological sources and the target variable is calculated, and the optimal meteorological characteristic data set is obtained by maximizing the correlation and minimizing the redundancy between NWP data.

Images